CA2012702A1 - Multispectral remote sensing of minerals using neural networks - Google Patents
Multispectral remote sensing of minerals using neural networksInfo
- Publication number
- CA2012702A1 CA2012702A1 CA002012702A CA2012702A CA2012702A1 CA 2012702 A1 CA2012702 A1 CA 2012702A1 CA 002012702 A CA002012702 A CA 002012702A CA 2012702 A CA2012702 A CA 2012702A CA 2012702 A1 CA2012702 A1 CA 2012702A1
- Authority
- CA
- Canada
- Prior art keywords
- data
- minerals
- remote sensing
- tape
- factors
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Abandoned
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
ABSTRACT OF THE DISCLOSURE
A method of remote sensing of minerals using neural networks to process data recorded on tape from an overflying aircraft. Vector quantization applies self-organizing feature maps to the clustering of spectral data from remotely sensed images. Normalized values of wavelengths of the spectral data are applied to a two-layer Kohonen self-organized feature map for a preselected number of classes of minerals.
A method of remote sensing of minerals using neural networks to process data recorded on tape from an overflying aircraft. Vector quantization applies self-organizing feature maps to the clustering of spectral data from remotely sensed images. Normalized values of wavelengths of the spectral data are applied to a two-layer Kohonen self-organized feature map for a preselected number of classes of minerals.
Description
2~27~2 MULTISPECTRAL REMOTE SENSING OF MINERALS VSING NEURAL
NETWORKS
BACKGROUND OF THE INVENTION
The present invention pertains to 5 multispectral processing, and more particularly, to multispectral processing using neural networks for remote sensing of minerals.
- Remote sensing of the Earth's surface from aircraft and from spacecraft using reflectance, 10 provides information not easily acquired by surface observations. Orbital imaging radar can be used in arid regions to provide subsurface data and in other regions radiometry can be used to identify minerals in the surface, and stressed vegetation can be used to 15 identify minerals in the substrate. Remote sensing imaging spectrometry consists of the acquisition of images in many narrow contiguous spectral bands throughout the visible and solar-reflected infrared spectral bands simultaneously. Imaging spectrometry 20 makes it possible to acquire data in hundreds of spectral bands simultaneously. These narrow spectral bands can be analyzed to determine the presence and types of mineral components. An overflying aircraft can acquire data from a plurality of areas by 25 recording a complete reflectance spectrum for each picture element or pixel in the image. The data associated with each pixel is removed to a work station where the data can be analyzed.
SUMMARY OF THE INVENT ON
The present invention senses the presence of mineral by using vector quantization which applies self-organizing feature maps to the clustering of spectral data from remotely sensed images. Data which is recorded on tape from an overflying aircraft is - ~,' ~ '' 2012 ~02 moved to a work station where the data is read into a computing system which calculates a normalized value for each of the discrete spectral wavelengths recorded on the tape. The normalized values of the wavelengths 5are applied to a two-layer Xohonen self-organized feature map for a preselected number of classes of minerals. The network is initialized using low random values of data, and forgetting factors and contraction factors from previous data are applied to the l0network. A training session is conducted until the segmentation process converges using the normalized values of the wavelengths. The converged process provides a listing of a preselected number of mineral classes which may be associated with a mineral being 15 sought. A sort and merge algorithm can be used to reduce the number of mineral classes if desired.
~RIEF_DESCRIPTION OF THE DRAWINGS
Figure 1 illustrates the method of collecting mineral data from an area on Earth by an overflying 20 aircraft.
Figure 2 illustrates typical reflectance curves of some common mineral compounds.
Figure 3 illustrates a self-organizing feature map having a plurality of discrete spectral 25 wavelengths as inputs and a number of mineral classes as outputs.
DESCRIPTION OF THE PREFERRED EMBODIMENT
Aircraft 10 flying over an area on the Earth E in search of certain minerals use a data recoFding 30 system 11 of the type disclosed in Figure 1.
Recording system 11 includes a plurality of lenses 16 - 18, a pair of mirrors 22, 23 and a dispersing element 24 to scan a plurality of areas (pixels) 28 on the Earth E to record a reflectance spectrum from each 35 area 28 on a corresponding line area 29 of a tape or -\ 20127~2 . .
film 30. The dispersing element 24 spreads the reflected light energy into a plurality of spectral bands along the line area 29. The recorded tape is then taken to a computer work station where the 5 multispectral data is read from each of the line areas 29 on the tape.
Figure 2 illustrates the general spectral response curves for some common minerals. The response curves are each shown as a continuous line 10 although they consist of a large number of closely-spaced, discrete vertical spectral lines with each line having a discrete reflectance amplitude and a discrete wavelength. Some spectral lines may be very close to the adjacent lines while there may be 15 gaps between other adjacent lines. When the wavelength portion of the graph is stretched (in a horizontally direction) the individual spectral lines can be more readily illustrated. Such spectral lines are well-known in the study of light and optics as a 20 means of identifying chemical elements.
At the computer work station the multispectral data associated with each of the pixels 28 (Fig. 1) is read from the tape 30 onto a computer disk. The multispectral data is normalized by adding 25 all of the individual spectral values for each pixel and dividing the total into each individual value at each spectral wavelength. The spectral values are applied to a two-layer Kohonen self-organizing feature Imap for a preselected number of classes as shown in Figure 3. The spectral input values are shown at the lower portion of Figure 3 and the output classes of minerals are shown at the upper portion of Figure 3.
The computer network is initialized using low random values of data; and forgetting factors and contraction factors previously developed on similar data such as - ~ 20127~2 :' ' ' -synthetic data or previously sensed data are applied.
A training session is conducted until the segmentation process converges.
The present method of u~ing self-organizing 5feature maps to cluster spectral data uses some of the background disclosed in a publication written by T.
Kohonen entitled ~Self-Organization And Associative Memory~, published by Spring-Verlog, Berlin, Germany 1984 and another article by F.H. Wu and K. Ganeson 10entitled ~An Algorithm Por Robust Vector Quantization Using A Neural-Net Model~, IEEE International Conference on Neural Networks Poster Session, San Diego, California July 1988.
The Kohonen neural-network clustering 15algorithm used operates in the following steps~
Step (1) Given a neural-network of size (J, I), where J is the number of vector patterns (neurons) and I the size of each vector pattern, initialize the weights Wij between input i 20and output j nodes to small random values.
Then let x (t); t = O,..., n - 1 represent the training vector sequence.
Step (2) Present a new input vector x (t) Step (3) Compute distance dj between the input vector 25pattern and each of the output vector nodes (j = O,...,J), dj = ~ (Xi(t) - Wij(t) )2 (3) i=O
-`~ 20127~2 Step ~43 Select the output vector node j* with minimum dj and adaptively modify its contents and its neighboring Euclidean distance vector patterns by Wij (t + 1) = Wij(t) + a (t)(Xi(t) - Wij(t)) (4) for j NEj*(t) and O < i < I - 1 NEj*(t) is a Euclidean distance neighborhood around the selected vector node which is decreased with time. The adaptation gain ;
term 0 C a(t) C 1 also decreases with time.
.i Step (5) Go to step (2) ~
:
If it appears that the output data is confusing due to the large number of mineral classes J
= 1, J = 2, etc. of Figure 3, then a sort and merge 15 algorithm can be used to consolidate the output into a smaller number of classes.
Although the best mode contemplated for carrying out the present invention has been herein shown and described, it will be apparent that 20 modification and variation may be made without departing from what is regarded to be the subject matter of the invention.
LBG:smb '.' '""''"''",.'''~
,'`', ' '' ',: - ,~
'`"' ;' '' `' ~''`
:, " ' .. ', '. ~:' . ,~
"'' ''" '"' :: ~ ' " ' '~ ': ' ,'`~ ~,'-,"',','','
NETWORKS
BACKGROUND OF THE INVENTION
The present invention pertains to 5 multispectral processing, and more particularly, to multispectral processing using neural networks for remote sensing of minerals.
- Remote sensing of the Earth's surface from aircraft and from spacecraft using reflectance, 10 provides information not easily acquired by surface observations. Orbital imaging radar can be used in arid regions to provide subsurface data and in other regions radiometry can be used to identify minerals in the surface, and stressed vegetation can be used to 15 identify minerals in the substrate. Remote sensing imaging spectrometry consists of the acquisition of images in many narrow contiguous spectral bands throughout the visible and solar-reflected infrared spectral bands simultaneously. Imaging spectrometry 20 makes it possible to acquire data in hundreds of spectral bands simultaneously. These narrow spectral bands can be analyzed to determine the presence and types of mineral components. An overflying aircraft can acquire data from a plurality of areas by 25 recording a complete reflectance spectrum for each picture element or pixel in the image. The data associated with each pixel is removed to a work station where the data can be analyzed.
SUMMARY OF THE INVENT ON
The present invention senses the presence of mineral by using vector quantization which applies self-organizing feature maps to the clustering of spectral data from remotely sensed images. Data which is recorded on tape from an overflying aircraft is - ~,' ~ '' 2012 ~02 moved to a work station where the data is read into a computing system which calculates a normalized value for each of the discrete spectral wavelengths recorded on the tape. The normalized values of the wavelengths 5are applied to a two-layer Xohonen self-organized feature map for a preselected number of classes of minerals. The network is initialized using low random values of data, and forgetting factors and contraction factors from previous data are applied to the l0network. A training session is conducted until the segmentation process converges using the normalized values of the wavelengths. The converged process provides a listing of a preselected number of mineral classes which may be associated with a mineral being 15 sought. A sort and merge algorithm can be used to reduce the number of mineral classes if desired.
~RIEF_DESCRIPTION OF THE DRAWINGS
Figure 1 illustrates the method of collecting mineral data from an area on Earth by an overflying 20 aircraft.
Figure 2 illustrates typical reflectance curves of some common mineral compounds.
Figure 3 illustrates a self-organizing feature map having a plurality of discrete spectral 25 wavelengths as inputs and a number of mineral classes as outputs.
DESCRIPTION OF THE PREFERRED EMBODIMENT
Aircraft 10 flying over an area on the Earth E in search of certain minerals use a data recoFding 30 system 11 of the type disclosed in Figure 1.
Recording system 11 includes a plurality of lenses 16 - 18, a pair of mirrors 22, 23 and a dispersing element 24 to scan a plurality of areas (pixels) 28 on the Earth E to record a reflectance spectrum from each 35 area 28 on a corresponding line area 29 of a tape or -\ 20127~2 . .
film 30. The dispersing element 24 spreads the reflected light energy into a plurality of spectral bands along the line area 29. The recorded tape is then taken to a computer work station where the 5 multispectral data is read from each of the line areas 29 on the tape.
Figure 2 illustrates the general spectral response curves for some common minerals. The response curves are each shown as a continuous line 10 although they consist of a large number of closely-spaced, discrete vertical spectral lines with each line having a discrete reflectance amplitude and a discrete wavelength. Some spectral lines may be very close to the adjacent lines while there may be 15 gaps between other adjacent lines. When the wavelength portion of the graph is stretched (in a horizontally direction) the individual spectral lines can be more readily illustrated. Such spectral lines are well-known in the study of light and optics as a 20 means of identifying chemical elements.
At the computer work station the multispectral data associated with each of the pixels 28 (Fig. 1) is read from the tape 30 onto a computer disk. The multispectral data is normalized by adding 25 all of the individual spectral values for each pixel and dividing the total into each individual value at each spectral wavelength. The spectral values are applied to a two-layer Kohonen self-organizing feature Imap for a preselected number of classes as shown in Figure 3. The spectral input values are shown at the lower portion of Figure 3 and the output classes of minerals are shown at the upper portion of Figure 3.
The computer network is initialized using low random values of data; and forgetting factors and contraction factors previously developed on similar data such as - ~ 20127~2 :' ' ' -synthetic data or previously sensed data are applied.
A training session is conducted until the segmentation process converges.
The present method of u~ing self-organizing 5feature maps to cluster spectral data uses some of the background disclosed in a publication written by T.
Kohonen entitled ~Self-Organization And Associative Memory~, published by Spring-Verlog, Berlin, Germany 1984 and another article by F.H. Wu and K. Ganeson 10entitled ~An Algorithm Por Robust Vector Quantization Using A Neural-Net Model~, IEEE International Conference on Neural Networks Poster Session, San Diego, California July 1988.
The Kohonen neural-network clustering 15algorithm used operates in the following steps~
Step (1) Given a neural-network of size (J, I), where J is the number of vector patterns (neurons) and I the size of each vector pattern, initialize the weights Wij between input i 20and output j nodes to small random values.
Then let x (t); t = O,..., n - 1 represent the training vector sequence.
Step (2) Present a new input vector x (t) Step (3) Compute distance dj between the input vector 25pattern and each of the output vector nodes (j = O,...,J), dj = ~ (Xi(t) - Wij(t) )2 (3) i=O
-`~ 20127~2 Step ~43 Select the output vector node j* with minimum dj and adaptively modify its contents and its neighboring Euclidean distance vector patterns by Wij (t + 1) = Wij(t) + a (t)(Xi(t) - Wij(t)) (4) for j NEj*(t) and O < i < I - 1 NEj*(t) is a Euclidean distance neighborhood around the selected vector node which is decreased with time. The adaptation gain ;
term 0 C a(t) C 1 also decreases with time.
.i Step (5) Go to step (2) ~
:
If it appears that the output data is confusing due to the large number of mineral classes J
= 1, J = 2, etc. of Figure 3, then a sort and merge 15 algorithm can be used to consolidate the output into a smaller number of classes.
Although the best mode contemplated for carrying out the present invention has been herein shown and described, it will be apparent that 20 modification and variation may be made without departing from what is regarded to be the subject matter of the invention.
LBG:smb '.' '""''"''",.'''~
,'`', ' '' ',: - ,~
'`"' ;' '' `' ~''`
:, " ' .. ', '. ~:' . ,~
"'' ''" '"' :: ~ ' " ' '~ ': ' ,'`~ ~,'-,"',','','
Claims (4)
1. A method of remote sensing of minerals using data recorded in pixels on tape in overflying aircraft which receives reflectance from elements on the Earth, where the data includes a plurality of narrow discrete spectral frequencies with each of the spectral frequencies having a recorded intensity, said sensing method including the steps of:
reading the multispectral data associated with each pixel from the tape into a work station;
calculating a normalized value for each of said discrete spectral frequencies;
using a neural network to separate the normalized multispectral data into a predetermined number of groups;
applying each of said normalized values to a two-layer Kohonen self-organizing feature map for a preselected number of classes of minerals; and conducting an unsupervised training session until a segmentation process converges.
reading the multispectral data associated with each pixel from the tape into a work station;
calculating a normalized value for each of said discrete spectral frequencies;
using a neural network to separate the normalized multispectral data into a predetermined number of groups;
applying each of said normalized values to a two-layer Kohonen self-organizing feature map for a preselected number of classes of minerals; and conducting an unsupervised training session until a segmentation process converges.
2. A method of remote sensing minerals as defined in claim 1 including the further step of developing a procedure for finding efficient startup network parameters.
3. A method of remote sensing as defined in claim 1 including the further step of initializing the neural network using low random values of data.
4. A method of remote sensing as defined in claim 1 including the further steps of developing forgetting factors and contraction factors using similar data and applying these factors to the collected multispectral data to achieve a sufficiently rapid learning procedure.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US33531189A | 1989-04-10 | 1989-04-10 | |
| US07/335,311 | 1989-04-10 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CA2012702A1 true CA2012702A1 (en) | 1990-10-10 |
Family
ID=23311240
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CA002012702A Abandoned CA2012702A1 (en) | 1989-04-10 | 1990-03-21 | Multispectral remote sensing of minerals using neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CA (1) | CA2012702A1 (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6321216B1 (en) * | 1996-12-02 | 2001-11-20 | Abb Patent Gmbh | Method for analysis and display of transient process events using Kohonen map |
| US11519781B2 (en) * | 2016-07-22 | 2022-12-06 | Nec Corporation | Image processing device, image processing method, and recording medium |
-
1990
- 1990-03-21 CA CA002012702A patent/CA2012702A1/en not_active Abandoned
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6321216B1 (en) * | 1996-12-02 | 2001-11-20 | Abb Patent Gmbh | Method for analysis and display of transient process events using Kohonen map |
| US11519781B2 (en) * | 2016-07-22 | 2022-12-06 | Nec Corporation | Image processing device, image processing method, and recording medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Gao et al. | NWPU-MOC: A benchmark for fine-grained multicategory object counting in aerial images | |
| Gao et al. | Spectral superresolution of multispectral imagery with joint sparse and low-rank learning | |
| Zheng et al. | Coupled convolutional neural network with adaptive response function learning for unsupervised hyperspectral super resolution | |
| Shawky et al. | Remote sensing image scene classification using CNN-MLP with data augmentation | |
| Bergado et al. | Recurrent multiresolution convolutional networks for VHR image classification | |
| Paoletti et al. | Capsule networks for hyperspectral image classification | |
| Romero et al. | Unsupervised deep feature extraction for remote sensing image classification | |
| Ouchra et al. | Satellite image classification methods and techniques: A survey | |
| Ghamisi et al. | Advances in hyperspectral image and signal processing: A comprehensive overview of the state of the art | |
| Zeggada et al. | A deep learning approach to UAV image multilabeling | |
| Li et al. | Integrating multilayer features of convolutional neural networks for remote sensing scene classification | |
| Liao et al. | Processing of multiresolution thermal hyperspectral and digital color data: Outcome of the 2014 IEEE GRSS data fusion contest | |
| Zhai et al. | Total variation regularized collaborative representation clustering with a locally adaptive dictionary for hyperspectral imagery | |
| Kerner et al. | Toward generalized change detection on planetary surfaces with convolutional autoencoders and transfer learning | |
| Tulczyjew et al. | Unsupervised feature learning using recurrent neural nets for segmenting hyperspectral images | |
| Li et al. | Structure-aware collaborative representation for hyperspectral image classification | |
| Ozkan et al. | Cloud detection from RGB color remote sensing images with deep pyramid networks | |
| Rao et al. | Overview of image processing | |
| Naseer et al. | Hybrid deep learning aerial framework for road scene objects segmentation and classification | |
| Paola et al. | Comparisons of neural networks to standard techniques for image classification and correlation | |
| Hidayat et al. | Grouping of image patterns using inceptionv3 for face shape classification | |
| KR940009990B1 (en) | Pattern Recognition System Using Neural Network | |
| Li et al. | Model-aware deep learning for the clustering of hyperspectral images with context preservation | |
| Chowdhury et al. | Neural network based dunal landform mapping from multispectral images using texture features | |
| CA2012702A1 (en) | Multispectral remote sensing of minerals using neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| FZDE | Discontinued |